The project consists in testing the algorithm on an environment that is assigned to us and analyze the data that it is working in that environment. In this project, we use Reinforcement Learning, which is a technique that allows the agent to learn a certain behavior depending on the environment submit to it with methods based on the concept of reward.
In my case, I was given the DQN algorithm to test on Assault environment of OpenAi Gym. The algorithm was provided to us using the ”stable baselines” library.
In the environment ”Assault-v0” the observation consists in the image, the k frames received or it can be the ram; instead in ”Assault-ram-v0” is executed the observation is the ram. The agent acts following the commands of ATARI 2600: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE, UPFIRE.
The reward starts from zero with each episode of the game and increases every time a spacecraft is hit, so the reward depends on the action done by the agent. When the agent dies, it returns to the initial state in which the reward is set to zero.
Our aim is to train the game using DQN algorithm, in particular, to modify hyperparameter in order to obtain quite significant results.
Use google colab to run the code, using the GPU. Google Colab